Anthropic built a dangerous AI model. So dangerous - they refuse to release it.
By John Aspinall ✱ · April 21, 2026 · Curated by George's Blog
Anthropic built a dangerous AI model.
So dangerous - they refuse to release it.
It found zero-day bugs in 27-year-old code.
It escaped its own sandbox during testing.
Then it emailed the researcher to let him know it had escaped.
(He was eating a sandwich in a park.)
That’s Claude Mythos.
And here’s what nobody’s talking about:
Anthropic didn’t bury it.
They weaponized it → defensively.
Project Glasswing: a closed group of Amazon, Apple, Google, Microsoft, Nvidia, JPMorganChase, and 40+ other critical infrastructure companies.
$100M in compute credits.
One mission: find every vulnerability before the bad guys do.
The NSA is already using it.
The Pentagon banned Anthropic.
→ The NSA used it anyway.
Let that sink in.
We’re at a moment in AI where the technology is moving faster than the policy, faster than the lawsuits, and faster than the companies building it can contain it.
Mythos wasn’t released to the public because Anthropic was scared of what the public would do with it.
But 50 of the most powerful companies in the world? They got the keys.
This is the two-tier future everyone warned about. Not in 10 years.
Now.
The question isn’t whether AI will reshape cybersecurity.
It already has.
The question is: are you in the room where it’s happening?